A flask-api (using gunicorn) is used as an inference api of a deep learning model.
This specific inference process is very cpu intensive (not using gpu yet).
What is the best practice of deploying it to a kubernetes cluster, based on these aspects:
should I create multiple pods handling requests using single gunicorn worker or less pods enabling gunicorn multiple workers? (node memory footprint)
since google provides to expose your deployment as a service using an external loadbalancer,
do I need a nginx web server on my flask-gunicorn stack?
creating multiple identical pods on the same node, is it more memory intensive than handling all these requests using multithreading on a single pod?
More smaller pods is generally better, provided you're staying under "thousands". It is easier for the cluster to place a pod that requires 1 CPU and 1 GB of RAM 16 times than it is to place a single pod that requires 16 CPU and 16 GB RAM once. You usually want multiple replicas for redundancy, to tolerate node failure, and for zero-downtime upgrades in any case.
If the Istio Ingress system works for you, you may not need separate a URL-routing layer (Nginx) inside your cluster. If you're okay with having direct access to your Gunicorn servers with no routing or filtering in front of that, directly pointing a LoadBalancer Service at them is a valid choice.
Running 16 copies of 1 application will generally need more memory than 1 copy with 16 threads; how much more depends on the application.
In particular, if you load your model into memory and the model itself is large, but your multi-threaded setup can share a single copy of it, 1 large pod could use significantly less memory than 16 small pods. If the model is COPYed directly into the Docker image and the application code mmap()s it then you'd probably get to share memory at the kernel layer.
If the model itself is small and most of the memory is used in the processing, it will still use "more" memory to have multiple pods, but it would just be the cost of your runtime system and HTTP service; it shouldn't substantially change the memory required per thread/task/pod if that isn't otherwise shared.
I'm in a high availability project which includes deployment of 2-node high availability cluster for hot replacement of services (applications) running on the cluster nodes. The applications have inbound and outbound tcp connections as well as process udp traffic (mainly for communicating with ntp server).
The problem is pretty standard until one needs to provide a hot migration of services to backup node with all the data stored in RAM. Applications are agnostic of backup mechanisms and it is highly undesirable to modify them.
As only approach to this problem, I've come off with a duplication approach assuming that both cluster nodes will run the same applications repeating calculations of each other. In case of failure the primary server the backup server will become a primary.
However, I have not found any ready solution for proxy which will have synchronous port mirroring. No existing proxy servers (haproxy, dante, 3proxy etc.) support such feature as far as I know. Have I missed something, or I should write a new one from scratch?
A rough sketch of the functionality can be found here:
p.s. I assume that it is possible to compare traffic from the two clones of the same application...
UPDATE: See the answer I've provided below for the solution I eventually got set up on AWS.
I'm currently experimenting with methods to implement a global load-balancing layer for my app servers on Digital Ocean and there's a few pieces I've yet to put together.
The Goal
Offer highly-available service to my users by routing all connections to the closest 'cluster' of servers in SFO, NYC, LON, and eventually Singapore.
Additionally, I would eventually like to automate the maintenance of this by writing a daemon that can monitor, scale, and heal any of the servers on the system. Or I'll combine various services to achieve the same automation goals. First I need to figure out how to do it manually.
The Stack
Ubuntu 14.04
Nginx 1.4.6
node.js
MongoDB from Compose.io (formerly MongoHQ)
Global Domain Breakdown
Once I rig everything up, my domain would look something like this:
**GLOBAL**
global-balancing-1.myapp.com
global-balancing-2.myapp.com
global-balancing-3.myapp.com
**NYC**
nyc-load-balancing-1.myapp.com
nyc-load-balancing-2.myapp.com
nyc-load-balancing-3.myapp.com
nyc-app-1.myapp.com
nyc-app-2.myapp.com
nyc-app-3.myapp.com
nyc-api-1.myapp.com
nyc-api-2.myapp.com
nyc-api-3.myapp.com
**SFO**
sfo-load-balancing-1.myapp.com
sfo-load-balancing-2.myapp.com
sfo-load-balancing-3.myapp.com
sfo-app-1.myapp.com
sfo-app-2.myapp.com
sfo-app-3.myapp.com
sfo-api-1.myapp.com
sfo-api-2.myapp.com
sfo-api-3.myapp.com
**LON**
lon-load-balancing-1.myapp.com
lon-load-balancing-2.myapp.com
lon-load-balancing-3.myapp.com
lon-app-1.myapp.com
lon-app-2.myapp.com
lon-app-3.myapp.com
lon-api-1.myapp.com
lon-api-2.myapp.com
lon-api-3.myapp.com
And then if there's any strain on any given layer, in any given region, I can just spin up a new droplet to help out: nyc-app-4.myapp.com, lon-load-balancing-5.myapp.com, etc…
Current Working Methodology
A (minimum) trio of global-balancing servers receive all traffic.
These servers are "DNS Round-Robin" balanced as illustrated in this
(frankly confusing) article: How To Configure DNS Round-Robin Load
Balancing.
Using the Nginx GeoIP
Module and
MaxMind GeoIP Data
the origin of any given request is determined down to the
$geoip_city_continent_code.
The global-balancing layer then routes the request to the least
connected server on the load-balancing layer of the appropriate
cluster: nyc-load-balancing-1, sfo-load-balancing-3,
lon-load-balancing-2, etc.. This layer is also a (minimum) trio of
droplets.
The regional load-balancing layer then routes the request to the
least connected server in the app or api layer: nyc-app-2,
sfo-api-1, lon-api-3, etc…
The details of the Nginx kung fu can be found in this tutorial:
Villiage Idiot: Setting up Nginx with GSLB/Reverse Proxy on
AWS. More general info about Nginx load-balancing is available
here
and
here.
Questions
Where do I put the global-balancing servers?
It strikes me as odd that I would put them either all in one place, or spread that layer out around the globe either. Say, for instance, I put them all in NYC. Then someone from France hits my domain. The request would go from France, to NYC, and then be routed back to LON. Or if I put one of each in SFO, NYC, and LON then isn't it still possible that a user from Toronto (Parkdale, represent) could send a request that ends up going to LON only to be routed back to NYC?
Do subsequent requests get routed to the same IP?
As in, if a user from Toronto sends a request that the global-balancing layer determines should be going to NYC, does the next request from that origin go directly to NYC, or is it still luck of the draw that it will hit the nearest global-balancing server (NYC in this case).
What about sessions?
I've configured Nginx to use the ip_hash; directive so it will direct the user to the same app or api endpoint (a node process, in my case) but how will global balancing affect this, if at all?
Any DNS Examples?
I'm not exactly a DNS expert (I'm currently trying to figure out why my CNAME records aren't resolving) but I'm a quick study when provided with a solid example. Has anyone gone through this process before and can provide a sample of what the DNS records look like for a successful setup?
What about SSL/TLS?
Would I need a certificate for every server, or just for the three global-balancing servers since that's the only public-facing gateway?
If you read this whole thing then reward yourself with a cupcake. Thanks in advance for any help.
The Goal: Offer highly-available service to my users by routing all connections to the closest 'cluster' of servers in SFO, NYC, LON, and eventually Singapore.
The global-balancing layer then routes the request to theleast
connected server...
If I'm reading your configuration correctly, you're actually proxying from your global balancers to the balancers at each region. This does not meet your goal of routing users to the nearest region.
There are three ways that I know of to get what you're looking for:
30x Redirect Your global balancers receive the HTTP request and then redirect it to a server group in or near the region it thinks the request is coming from, based on IP address. This sounds like what you were trying to set up. This method has side effects for some applications, and also increases the time it takes for a user to get data since you're adding a ton of overhead. This only makes sense if the resources you're redirecting to are very large, and the local regional cluster will be able to serve much more efficiently.
Anycast (taking advantage of BGP routing) This is what the big players like Akamai use for their CDN. Basically, there are multiple servers out on the internet with the exact same routable IP address. Suppose I have servers in several regions, and they have the IP address of 192.0.2.1. If I'm in the US and try to connect to 192.0.2.1, and someone is in Europe that tries to connect to 192.0.2.1, it's likely that we'll be routed to the nearest server. This uses the internet's own routing to find the best path (based on network conditions) for the traffic. Unfortunately, you can't just use this method. You need your own AS number, and physical hardware. If you find a VPS provider that lets you have a chunk of their Anycast block, let me know!
Geo-DNS There are some DNS providers that provide a service often marketed as "Geo-DNS". They have a bunch of DNS servers hosted on anycast addresses which can route traffic to your nearest servers. If a client queries a European DNS server, it should return the address for your European region servers, vs. some in other regions. There are many variations on the Geo DNS services. Others simply maintain a geo-IP database and return the server for the region they think is closer, just like the redirect method but for DNS before the HTTP request is ever made. This is usually the good option, for price and ease of use.
Do subsequent requests get routed to the same IP?
Many load balancers have a "stickiness" option that says requests from the same network address should be routed to the same end server (provided that end server is still up and running).
What about sessions?
This is exactly why you would want that stickiness. When it comes to session data, you are going to have to find a way to keep all your servers up-to-date. Realistically, this isn't always guaranteed. How you handle it depends on your application. Can you keep a Redis instance or whatever out there for all your servers to reliably hit from around the world? Do you really need that session data in every region? Or can you have your main application servers dealing with session data in one location?
Any DNS Examples?
Post separate questions for these. Everyone's "successful setup" looks differently.
What about SSL/TLS?
If you're proxying data, only your global balancers need to handle HTTPS. If you're redirecting, then all the servers need to handle it.
A Working Solution
I've had a wild ride over the past few months figuring out the whole Global-HA setup. Tonnes of fun and I've finally settled with a rig that works very well, and is nothing like the one outlined in the above question.
I still plan on writing this up in tutorial form, but time is scarce as I head into the final sprint to get my app launched early next year, so here's a quick outline of the working rig I ended up with.
Overview
I ended up moving my entire deployment to AWS. I love Digital Ocean, but the frank reality is that AWS is light years ahead of them (and everyone, really) when it comes to the services offered under one roof. My monthly expenses went up slightly, but once I was done tweaking and streamlining I ended up with a solution that costs about $75/month per region for the most basic deployment (2 instances behind an ELB). And a new region can be spun up and deployed within about 30 minutes.
Global Balancing
I quickly found out (thanks to #Brad's answer above) that trying to spin up my own global balancing DNS layer is insane. It was a hell of a lot of fun figuring out how a layer like this works, but short of getting on a plane and scraping my knuckles installing millions of dollars worth of equipment around the world, it was not going to be possible to roll my own.
When I finally figured out what I was looking for, I found my new best friend: AWS Route 53. It offers a robust DNS network with about 50-odd nodes globally and the ability to do some really cool routing tricks like location-based routing, latency-based routing (which is kinda awesome), and AWS Alias records that 'automagically' route traffic to other AWS Services you'll be using (Like ELB for load balancing).
I ended up using latency-based routing that directs the global traffic to the closest regional Elastic Load Balancer, which has an Auto-Scaling Group attached to it in any given region.
I'll leave it up to you to do your homework on the other providers: www.f5.com, www.dyn.com, www.akamai.com, www.dnsmadeeasy.com. Depending on your needs, there may be a better solution for you, but this works very well for me.
Content Delivery Network
Route 53 integrates with AWS Cloudfront very nicely. I setup an S3 bucket that I'm using to store all the static media files that my users will upload, and I've configured a Cloudfront distribution to source from my media.myapp.com S3 bucket. There are other CDN providers, so do your shopping. But Cloudfront gets pretty good reviews and it's a snap to setup.
Load Balancing & SSL Termination
I'm currently using AWS Elastic Load Balancer to balance the load across my application instances, which live in an Auto-Scaling Group. The request is first received by ELB, at which point SSL is terminated and the request is passed through to an instance in the Auto-Scaling Group.
NOTE: One giant caveat for ELB is that, somewhat ironically, it doesn't handle massive spikes very well. It can take up to 15 minutes for an ELB to trigger a scale-up event for itself, creating 500/timeouts in the meantime. A steady, constant increase in traffic is supposedly handled quite well, but if you get hit with a spike it can fail you. If you know you're going to get hit, you can 'call ahead' and AWS will warm up your ELB for you, which is pretty ridiculous and anti-pattern to the essence of AWS, but I imaging they're either working on it, or ignoring it because it's not really that big of a problem. You can always spin up your own HAProxy or Nginx load-balancing layer if ELB doesn't work for you.
Auto-Scaling Group
Each region has an ASG which is programmed to scale when the load passes a certain metric:
IF CPU > 90% FOR 5 MINUTES: SCALEUP
IF CPU < 70% FOR 5 MINUTES: SCALEDN
I haven't yet put the ELB/ASG combo through its paces. That's a little way down my To-Do list, but I do know that there are many others using this setup and it doesn't seem to have any major performance issues.
The config for an Auto-Scaling Group is a little convoluted in my opinion. It's actually a three-step process:
Create an AMI configured to your liking.
Create a Launch Configuration that uses the AMI you've created.
Create an Auto-Scaling Group that uses the Launch Configuration you've created to determine what AMI and instance type to launch for any given SCALEUP event.
To handle config and app deployment when any instance launches, you use the "User Data" field to input a script that will run once any given instance launches. This is possibly the worst nomenclature in the history of time. How "User Data" describes a startup script only the author knows. Anyhow, that's where you stick the script that handles all your apt-gets, mkdirs, git clones, etc.
Instances & Internal Balancing
I've also added an additional 'internal balancing layer' using Nginx that allows me to 'flat-pack' all my Node.js apps (app.myapp.com, api.myapp.com, mobile.myapp.com, www.myapp.com, etc.myapp.com) on every instance. When an instance receives a request passed to it from ELB, Nginx handles routing the request to the correct Node.js port for any given application. Sort of like a poor-mans containerization. This has the added benefit that any time one of my apps needs to talk to the other (like when app. needs to send a request to api.) it's done via localhost:XXXX rather than having to go out across the AWS network, or the internet itself.
This setup also maximizes usage of my resources by eliminating any idle infrastructure if the app layer it hosts happens to be receiving light traffic. It also obviates the need to have and ELB/ASG combo for every app, saving more cash.
There's no gotchas or caveats that I've run into using this sort of setup, but there is one work-around that needs to be in place with regard to health-checking (see below).
There's also a nice benefit in that all instances have an IAM role which means that your AWS creds are 'baked in' to each instance upon birth and accessible via your ENV vars. And AWS 'automagically' rotates your creds for you. Very secure, very cool.
Health Checks
If you go the route of the above setup, flat-packing all your apps on one box and running an internal load-balancer, then you need to create a little utility to handle the ELB Health Checks. What I did was create an additional app called ping.myapp.com. And then I configured my ELB Health Checks to send any health checks to the port that my ping app is running on, like so:
Ping Protocol: HTTP
Ping Port: XXXX
Ping Path: /ping
This sends all health checks to my little ping helper, which in turn hits localhost:XXXX/ping on all the apps residing on the instance. If they all return a 200 response, my ping app then returns a 200 response to the ELB health check and the instances gets to live for another 30 seconds.
NOTE: Do not use Auto-Scaling Health Checks if you're using an ELB. Use the ELB health checks. It's kinda confusing, I thought they were the same thing, they're not. You have the option to enable one or the other. Go with ELB.
The Data Layer
One thing that is glaringly absent from my setup is the data layer. I use Compose.io as my managed data-layer provider and I deploy on AWS so I get very low latency between my app layers and my data layer. I've done some prelim investigation on how I would roll my data layer out globally and found that it's very complex — and very expensive — so I've kicked it down my list as a problem that doesn't yet need to be solved. Worst case is that I'll be running my data layer in US-East only and beefing up the hardware. This isn't the worst thing in the world since my API is strictly JSON data on the wire so the average response is relatively tiny. But I can see this becoming a bottleneck at very large, global scale — if I ever get there. If anyone has any input on this layer I'd love to hear what you have to say.
Ta-Da!
Global High Availability On A Beer Budget. Only took me 6 months to figure it out.
Love to hear any input or ideas from anyone that happens to read this.
You can use Anycast for your webservice for free if using Cloudflare free plan.
Digital Ocean now supports Load Balancing of servers itself. It is extremely easy to set up and works great! Saves you having to add in unnecessary components such as nginx (if you only want to use for load balancing).
We were having issues using SSL file uploads with nginx on a digital ocean server, however since the Digital Ocean update, we have removed nginx and now use Digital Ocean's load balancing feature and it works just as we need it to!
I need to set up a tracking server that will only serve 1x1 pixels and log all requests.
I initially thought of using Amazon's S3 or CloudFront but their costs are prohibitively high for me. I need to serve 500M pixels a day, and S3 charges $0.4 per 1M GET requests, so even without the data transfer costs I'm at $6,000/month.
I am considering setting up nginx or lighttpd on an EC2 instance. What performance should I expect with those two (e.g. per one large EC2 instance)? Are there better free products for this task?
Nginx is indeed a good candidate for this and already has built in support for empty GIFs (see http://wiki.nginx.org/HttpEmptyGifModule).
Disk I/O will probably be the biggest issue for this server because of the access logging. The only way to figure out the performance of the different EC2 instances is to test them.
If one EC2 instance does not offer the performance you need, or if you need any redundancy for this service, you should also look into using a load balancer (either an AWS Elastic Load Balancer or your own custom one).
You could also set up multiple smaller servers in different geographical regions and use DNS latency based routing to route requests to them (use either AWS Route 53 latency based routing or another DNS solution). This would significantly reduce the connection time to your server and would distribute the load across several data centers.
At one of our client's site we have the following topology of BizTalk 2006 in production environment:
2 BizTalk Runtime servers
1 SQL Server with MsgBox and TrackingDB on it.
One of the runtime servers are dedicated physical server and the another is virtual server, though both have 4 CPUs and 8Gb of memory.
The physical server CPU is hardly in use while the virtual server is always on 50% - 60% of CPU usage.
Is there any way to configure the BizTalk load balancing algorithm in such a way that it would load on one server more than on the other?
Yes and no is the accurate answer :-)
Given one host, with one instance on each server, BizTalk would apply a simple round-robin approach to load balancing; as a service (orchestration or pipeline) is tied to a host the answer would be that for a single service it is not possible to assign more "weight" for a particular server -load would be spread evenly between the instances.
However, if you have multiple services, and it makes sense, you could distribute some of them exclusively to the physical box while some to both; this would mean that the physical box would take more load compare to the virtual one.
Obviously - on a two servers configuration - it does mean you lose redundancy though, a decision I would have not taken lightly so I would recommend against it.
(having said that - you say you have one SQL box, so perhaps redundancy is not a major issue?)
Also, if you're doing HTTP receives, check the donfiguration of your web load balancer. It maybe preferring one server over the other based on its configuration.
Another suggestion would be to disable the host instances on the VM and run primarily off of the physicaL box. If the physical box goes down, set up a MOM alert or such to kick off a script to start up the VMs host instances. We do something like this for FTP receives in our cluster.